Dr Junhao Liang awarded the Tom Lupton Prize
Well done to Dr Junhao Liang who graduated with his PhD and awarded the Tom Lupton Prize for his paper - Tabular Image: a method to convert tabular data to images for convolutional neural networks.
A PhD researcher from the Centre for Decision Research in the Analytics, Technology and Operations Department has been awarded the Tom Lupton Prize on the same day he graduated. The prize recognises his paper "Tabular Image: a method to convert tabular data to images for convolutional neural networks", co-authored with his supervisors Dr Xingjie Wei and Professor Barbara Summers, and published in the leading journal Annals of Operations Research.
Credit scoring sits at the heart of retail banking. Lenders use statistical models to estimate how likely an applicant is to default, and at the scale of a modern loan portfolio even a fraction of a percentage point of accuracy translates into very large sums of money saved or lost. Over the past decade, deep learning has produced extraordinary results in computer vision and natural language processing. Yet in credit scoring it has consistently underperformed. Benchmarking studies have repeatedly found that gradient boosting methods such as XGBoost match or beat neural networks on credit data.
The reason is structural rather than a failure of effort. Credit data is tabular: rows of applicants, columns of mixed numerical, ordinal and categorical features, frequently with missing values and many uninformative variables. Neural networks, and particularly convolutional neural networks (CNNs), are built to exploit the spatial structure of images, where neighbouring pixels are meaningfully related. A spreadsheet has no such structure. Researchers have therefore been confined to architectures that accept one-dimensional input, leaving the most powerful and best-developed part of the deep learning toolkit off the table.
The proposed method, Tabular Image, converts each individual record into a compact greyscale image that a state-of-the-art two-dimensional CNN can read. The elegance of the approach lies in how the image is constructed. Tabular Image builds the image around two measures that credit analysts have used for decades: weight of evidence and information value. Weight of evidence converts categorical values into numbers that reflect how strongly each value separates defaulting from non-defaulting applicants. Information value then quantifies how much each feature contributes overall, and this determines how many pixels that feature is allocated in the image. More informative features occupy more of the picture. Features are then arranged so that correlated variables sit in adjacent blocks, meaning that spatial proximity in the image carries genuine information — exactly the property a convolutional network is designed to exploit. The result is a compact, information-dense image rather than the sparse, largely blank images produced by earlier one-hot encoding approaches.
Tested across three benchmark credit datasets, the approach delivers state-of-the-art predictive performance. The transformation also has an interpretive benefit: defaulting and non-defaulting applicants produce visibly distinct images, offering credit managers a more intuitive route into complex data. The framework is designed to generalise to tabular datasets in other domains.
The paper is published open access and can be read in full at: Tabular Image: a method to convert tabular data to images for convolutional neural networks | Annals of Operations Research | Springer Nature Link
